By the end of this chapter you can write a voice file precise enough that Claude sounds like you, not like generic AI, and you can spot the tells that give away machine-written text.
Your constitution had a short how-you-sound section. This chapter deepens it into a real voice file, because the single most common complaint about AI writing is that it all sounds the same. Bland, padded, weirdly cheerful, instantly recognisable as machine-made. You are going to learn why, and how to make yours not do that.
There are two halves to sounding like you. The first is removing the slop: the puffery, the empty enthusiasm, the patterns that scream this was generated. The second, which almost everyone forgets, is adding the soul: the specific, slightly odd, human things that make writing yours. Strip the slop and stop there and you get clean, lifeless prose, which is its own kind of tell. You need both.
AVATAR OPENER · ~90s
Watch: making the same paragraph sound generic, then unmistakably like a person
HeyGen avatar · generated, consistent presenter
The reason AI writing has a sameness is the same reason its answers are generic without context. Left alone, it reaches for the most statistically likely words, and the most likely words are the ones everyone uses. A voice file is how you pull it off that default. It is a sample of you plus a list of your tells to avoid, and it is the difference between writing that could be anyone\u2019s and writing that could only be yours.
Generic AI voice
Puffery, empty enthusiasm, rule-of-three filler, the same upbeat register everyone gets. Recognisable on sight.
Your rhythm, your blunt bits, the things only you say and the things you would never say. A sample to match.
Result: reads like a person. Specifically, like you.
Sounding like you is two jobs: cut the slop, then add the soul.
Strip the slop and stop, and you get clean, lifeless prose, which is its own tell. Remove the slop AND add the soul.
THE TELLS TO CUT
Once you can name the tells, you see them everywhere, and so can the AI when you tell it to. These four are the most common giveaways of machine-written text.
Everything in tidy lists of three. Real people are not that neat.
It is not just X, it is Y
The signature AI sentence shape. Once you see it, you cannot unsee it.
Relentless cheer
Everything is exciting and great. Real voice has dry bits, doubt, edges.
SEE IT
Here is a bio paragraph, generated and left alone, then rewritten to sound like an actual person. Drag the handle. Notice it is not just shorter, it has texture: a real detail, a dry aside, something a machine would not volunteer.
Generic (slop, no soul)I am a passionate and dedicated marketing professional with a proven track record of delivering impactful, results-driven campaigns. I love helping businesses unlock their full potential and elevate their brand to the next level.
Cut the slop, added the soulI do marketing for small food brands, mostly the ones too stubborn to sound like everyone else. Twelve years of it. I am good at the part most people skip, which is figuring out what is actually true about a product before writing a word about it.
could be anyonecould only be her
The voice file is what makes that rewrite repeatable instead of a one-off. You write down how you sound and what you would never write, you feed it a real sample of your own writing so it learns your patterns, and from then on you can ask the AI to write in your voice and check its own drafts against your tells. Here is a starter you can keep.
Your voice file (copy and fill)
# How I sound
## In three words
[e.g. plain, dry, warm]
## My rhythm
[Short sentences or long? Do you start with the point or build to it?
Paste two or three sentences you actually wrote and like.]
## Words and moves I use
[The phrases, the kind of detail, the bluntness or warmth that is yours.]
## Words and moves I would NEVER use
[Your banned list. e.g. passionate, journey, unlock, elevate, exclamation
marks, the "it is not just X, it is Y" shape, tidy lists of three.]
## When checking my writing
Flag anything from the NEVER list. Flag any sentence that could have been
written about anyone. Keep the bits that sound like a real person, even
the slightly odd ones.
Fill in the voice file with your own real patterns. Crucially, paste in two or three sentences you actually wrote and like, so it has a true sample of you. Then ask Claude to write one real thing in your voice using the file, and separately to audit a piece of generic text against your never-list. Judge whether the output sounds like a person, specifically you.
Right if the voiced output keeps at least one specific, human detail and avoids every item on your never-list, and a piece that knows you would say it sounds like you.
Show the worked solution
The make-or-break ingredient is the real sample. A voice file that only lists adjectives (plain, warm) gives the AI a description; a voice file with two sentences you actually wrote gives it a target to match, and matching beats describing every time. Test both halves. First the cutting: feed it the puffed-up marketing bio above and your never-list, and it should flag passionate, proven track record, unlock, elevate, results-driven and the relentless cheer. That is the easy half, and it works fast. Now the harder half, the soul: ask it to write your bio in your voice. If the result is clean but bland ("I help food brands with marketing. I have twelve years of experience.") the slop is gone but so is the person. Push it: "keep it accurate but put back the texture, a real detail and a dry aside, the way my sample sentences sound." Now you get the stubborn-brands line, the part-most-people-skip line. That texture is the soul, and it only comes back because you gave it a real sample to reach for. Clean is not the goal. You is the goal.
WATCH FOR
✗Your voice file is a list of adjectives. Adjectives describe. A real sample of your writing gives it something to match. Paste one in.
✗You strip the slop and stop. Clean but lifeless is its own tell. Add the soul: the specific, human, slightly odd bits.
✗You cannot name your own tells. Borrow the common ones (puffery, rule of three, relentless cheer) and add what bugs you personally.
✗You accept "sounds professional". Professional is often just polished slop. The bar is sounds like a real person, specifically you.
WHAT YOU LEARNED
The takeaways
AI writing has a sameness because, left alone, it reaches for the most likely words, which are the ones everyone uses.
Sounding like you is two jobs: cut the slop (puffery, rule of three, the X-not-Y shape, relentless cheer) and add the soul (your specific, human texture).
A voice file is a sample of you plus your never-list. The real sample is what lets the AI match you instead of describing you.
Clean is not the goal. Lifeless clean is its own tell. The bar is reads like a real person, specifically you.
Your project · step sixteen
Add your voice file to your thread project, alongside the constitution. Now everything it writes can be checked against how you actually sound and what you would never say. Your project does not just know what you do, it knows how you sound doing it. Next chapter connects it to the outside world, so it can pull in live information instead of guessing from frozen memory.
Everyone has access to the same model. The thing that makes your output yours is the part that is made of you. Write it down once, and it sounds like you forever.